A Hybrid Perturbed Secret Sharing Federated Learning Framework
Yinglin Feng, Shuhong Chen, Zidong Cheng, Weifeng Su, Tian Wang · 2024
In recent years, the problem of privacy leakage in federated learning has received much attention. To address this issue, researchers have applied privacy preserving techniques to federated learning. However, differential privacy adding too much noise can affect the generalization of the model and the availability of data. Secret sharing may also lead to data leakage if attacked. In order to tackle the problems of excessive noise and data security, this paper proposes a hybrid perturbation secret sharing federated learning framework (HPS-FL). It optimizes local difference privacy based on mean statistics and mitigates the effect of excessive noise by changing the perturbation boundary and reducing the variance. It also introduces secret sharing technology, which further increases the security of parameter transmission by adding a mask before secret sharing. Experimental results on the MINIST dataset demonstrate that HPS-FL outperforms other methods in terms of model accuracy, providing reliable support for privacy preserving.